aitextgen
aitextgen copied to clipboard
Glitchy initial generations (missing text, weird punctuation)
The initial generations of some texts tend to have missing text, weird punctuation not present in the original dataset. This could be a function of tokenization.
Although the generation changes in 0.5.0 mitigate this, I should probably verify it's not another tokenization issue that could cause breakage.
Possibly fixed by https://github.com/minimaxir/aitextgen/commit/62dce28249dd019a4be61ce94961fd4597403562 as recent generations were not correctly reading the "beginning."